Events Calendar

Thesis Proposal - Ryan Julian

Title: The Adaptation Base Case: Understanding the Challenge of Continual Robot Learning
Date/Time: Thursday, September 19th 12pm
Location: RTH 406
Candidate: Ryan Julian
Committee: Prof. Gaurav Sukhatme (adviser), Prof. Joseph Lim, Prof. Heather Culbertson, Prof. Stefanos Nikolaidis, Prof. SK Gupta, Dr. Karol Hausman

Abstract:
Much of the promise of reinforcement learning (RL) for robotics is predicated on the idea of hands-off continual improvement: that these systems will be able to use machine learning to improve their performance after deployment. Without this property, RL does not compare very favorably to hand-engineered robotics. The research community has successfully shown that RL can train agents which are at least as good, or better than, hand-engineered controllers after a single large-scale up-front training process. Furthermore, multi-task and meta-learning has research shown that we can learn controllers which adapt to new tasks, by reusing data and models from related tasks. What is not well-understood is whether we can make this adaptation process continual. The overall schematic off-policy multi-task RL algorithms suggests these might make good continual learners, but we don't if know that's actually the case. In this presentation, I'll review the recent history of adaptive robot learning research, and enumerate the most important unanswered questions which prevent us from designing continual multi-task learners. I'll then outline a research agenda which will answer those questions, to provide a road map to continual multi-task learning for robotics.

Location: 406

Audiences: Everyone Is Invited

Contact: Lizsl De Leon